AI-enabled systems must sustain dependable behavior as their operating conditions, data distributions, model capabilities, and governance constraints evolve. Traditional software self-adaptation relies on feedback-loop architectures; however, contemporary AI deployments require additional loop structures: continuous evaluation and retraining, runtime safeguarding and policy enforcement, edge/cloud placement control, and goal-directed distributed coordination among services and agents. This paper conducts a systematic, cross-domain literature review of feedback-loop mechanisms relevant to AI-enabled systems. Following a structured protocol (search, selection, extraction, and synthesis), we analyze 52 primary studies identified from 2,365 candidates. We synthesize a taxonomy-driven catalog of reusable control-loop patterns. The results are organized into three main contributions. First, we present a taxonomy of adaptive systems’ design decisions that govern how feedback loops are structured and operated. Second, we offer a catalog of AI-enabled systems design patterns that embed feedback loops into layered, pipeline, microservice, safeguard, and data-processing architectures. Third, we provide a catalog of deployment patterns that address edge/cloud control, federated learning and inference, inference cascades, privacy-preserving adaptation, safety governance, and online personalization. The findings highlight recurring multi-loop architectures spanning runtime, lifecycle, and governance concerns. The study concludes with cross-cutting insights on multi-loop stacks, assurance and stability, and open research challenges for learning-enabled control.